Prediction of Thermal Decomposition Temperatures Using Statistical Methods

We create and evaluate computational models for the prediction of onset thermal decomposition temperatures for energetic materials using machine learning techniques. Our models are trained and tested on published differential scanning calorimetry data consisting of decomposition temperatures and molecular formulas. The chemical information is encoded using cheminformatics molecular representations. Due to the high dimensionality of the descriptor space, we analyze the utility of explicit dimension reduction and regularization techniques. We contrast the performance of linear and nonlinear regression methods and discuss the dependence of prediction quality with dataset splitting into training and test sets. We also identify outliers and examine the effect of outlier removal on model performance. For nitro, nitrate, and nitramine compounds, we obtain statistical models with root mean square errors close to the estimated noise in the experimental decomposition temperatures of the underlying dataset.